[Flang] 178.galgel/spec2000: Investigate inlinling of matmul and mutmul(transpose,..) intrinsics

[kiranchandramohan]

Investigate whether inlining of matmul and matmul(transpose,...) intrinsics provides better performance.

@Radu2k (https://github.com/radu2k) is investigating this.

[llvmbot]

@llvm/issue-subscribers-flang-ir

Details

Investigate whether inlining of `matmul` and `matmul(transpose,...)` intrinsics provides better performance.

@Radu2k (https://github.com/radu2k) is investigating this.

[Radu2k]

We observed that inlining the matmul calls within SPEC2000 does not contribute to any significant performance improvement when HLFIR flag is enabled.

However, it's noteworthy that a speed increase can be achieved by inlining matrix times vector calls. This method offers a slight performance boost, albeit by a mere 0.2 seconds. Given the minimal impact of this optimisation, there is no point in inlining only these calls for such a small improvement.

Flang-new

Inline type	NO HLFIR	HLFIR
No inlining	24.381s	20.695s
Matrix-times-matrix inlined	24.456s	20.915s
Matrix-times-vector inlined	23.186s	20.441s
Matrix-times-matrix and matrix-times-vector inlined	23.518s	20.757s
Matrix-times-matrix, matrix-times-vector, and transposed-matrix-times-matrix inlined	20.953s	21.006s

Gfortran

Inline type	Result
No inlining	14.025s
Matrix-times-matrix inlined	15.673s
Matrix-times-vector inlined	13.584s
Matrix-times-matrix and matrix-times-vector inlined	15.052s
Matrix-times-matrix, matrix-times-vector, and transposed-matrix-times-matrix inlined	19.525s

[tarunprabhu]

Closing this since the investigation has revealed that inlining does not produce any significant performance advantage.

Feel free to reopen if there is more that needs to be done.